feat(harvey-lab): switch judge from Gemini to Claude Sonnet#3
Merged
Merged
Conversation
Replace the hardcoded Gemini judge with Claude Sonnet (claude-sonnet-4-6) to match Harvey LAB's original benchmark default judge model. Changes across all 1,251 tasks: - evaluate.py: call_gemini() → call_judge() using Anthropic API - task.toml: GEMINI_API_KEY → ANTHROPIC_API_KEY in [verifier.env] - Dockerfile: google-genai → anthropic pip dependency Matches the converter change in benchflow-ai/benchflow PR #264.
Author
🤖 Devin AI EngineerI'll be helping with this pull request! Here's what you should know: ✅ I will automatically:
⚙️ Control Options:
|
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Updates all 1,251 Harvey LAB pre-converted tasks to use Claude Sonnet (
claude-sonnet-4-6) as the LLM judge instead of Gemini (gemini-3.1-flash-lite-preview). This matches the original Harvey LAB benchmark's default judge model (per eval-strategies.md).\n\nThree files changed per task (3,753 files total):\n-tests/evaluate.py:call_gemini()→call_judge()using Anthropic API withclaude-sonnet-4-6\n-task.toml:GEMINI_API_KEY→ANTHROPIC_API_KEYin[verifier.env]\n-environment/Dockerfile:google-genai→anthropicpip dependency\n\nCompanion to benchflow-ai/benchflow#264 which updates the converter.\n\n## Review & Testing Checklist for Human\n\n- [ ] Run a single task locally to verify the Claude judge produces a valid reward:\nbash\n export ANTHROPIC_API_KEY=...\n bench eval create \\\n --source-repo benchflow-ai/benchmarks \\\n --source-path datasets/harvey-lab/tasks/corporate-ma-analyze-cim-deal-teaser-scenario-01 \\\n -a gemini -m gemini-3.1-flash-lite-preview -e docker\n\n- [ ] VerifyANTHROPIC_API_KEY(notGEMINI_API_KEY) appears in task.toml[verifier.env]\n- [ ] Spot-check that evaluate.py containsimport anthropicandclaude-sonnet-4-6, notgoogle-genai\n\n### Notes\n\n- All 1,251 evaluate.py files were identical (same md5), so a single template replacement was safe\n- All 1,251 Dockerfiles were also identical — sed replacement was used forgoogle-genai→anthropic\n- Agent-side behavior is unchanged — only the verifier (judge) is affected\n- Users now needANTHROPIC_API_KEYexported instead ofGEMINI_API_KEYfor the verifier"Link to Devin session: https://app.devin.ai/sessions/6b1fbe50a16b4fd497ed4a26cc1d4b6c
Requested by: @xdotli